Beating AI News Flash: New developments have emerged in the data breach incident at Fireworks, an AI inference and model training platform. Notion has confirmed that hackers obtained real chat records from some Notion AI users, including questions posed by users and AI-generated responses. Fewer than 1% of all workspaces were affected.
These chat records came from Notion's Insights feature. Notion sends user conversations with AI to Fireworks for analysis to help enterprise administrators understand how employees use AI.
Notion stated that Fireworks had contractually promised not to retain this data. However, an investigation found that Fireworks had actually stored some chat content, which was ultimately obtained by hackers.
On November 2, Fireworks initially notified Notion that only fictional data used for testing had been affected. Notion immediately cut off its connection with Fireworks on November 3. It was not until November 8 that Fireworks further confirmed that real customer data had also been accessed.
Notion said that the vast majority of leaked chat records did not include user names, company names, or workspace information, but the conversations themselves could still contain sensitive content such as trade secrets and passwords. The company has notified affected customers and recommends reviewing chat records and changing any passwords and API keys that may have been compromised.
Currently, Notion has disabled the Insights feature and no longer provides open-source AI models through Fireworks. The company is checking whether other AI service providers truly comply with their commitments not to retain customer data.
